This AI system can diagnose sepsis with 99% accuracy before it becomes life-threatening
Researchers fed medical data from patients at home, in the ambulance and in the emergency room to build to a machine learning model and built a nearly perfect diagnostic tool that could help doctors save lives
Using medical data from sick patients while they’re still at home, traveling in an ambulance and receiving care in the emergency room, researchers at Northeastern have used artificial intelligence to build a tool that predicts life-threatening septic shock with 99% accuracy.
Sepsis, an extreme immune system response to infection in the body, is the cause of death for one in three people who die in a hospital.
“If sepsis is diagnosed in the emergency room, probably the best-case scenario is to pray because the survival rate is extremely low,” says Sergey Aitan, teaching professor in Northeastern University’s Multidisciplinary Engineering Graduate Programs on the Oakland campus and a lead investigator in the project. “Our system is like an immediate second opinion, which is practically impossible to do in emergency settings with physical doctors.”
Researchers used patient data like severe fever, chills, breathing difficulty, skin discoloration, fatigue and confusion captured when a patient is at home, on the way to the hospital and in the emergency room to train a machine learning model to predict sepsis.
Early onset of sepsis is hard to identify, Aitan says, because symptoms are subtle and overlap with other conditions. The research was published in the journal of the Multidisciplinary Digital Publishing Institute. Co-investigators include Rolando Herrero, director of Northeastern’s master’s programs in cyber-physical systems and telecommunications networks, assistant professor of artificial intelligence Abdolreza Mosaddegh and adjunct engineering faculty Haitham Tayyar and graduate students Ebunoluwa Adebesin, Sai Pranavi Jeedigunta and Hangyeol Kim.
“There is no other research that basically takes into account those three stages,” says Herrero. “Our students collaborated to create this innovative AI model that enables doctors to detect sepsis very early in the game.”
The three-stage approach improves accuracy, says Aitan. The machine learning model’s predictions are correct 82% of the time when it only has a patient’s descriptions of their symptoms. When ambulance vital signs are added, predictions are correct 99% of the time and emergency room test results push accuracy up even further.
In practice, Aitan says, emergency room physicians would type or speak into their phones to enter patient information in any language and the system predicts the likelihood that the patient will develop sepsis.
Researchers obtained sepsis patient data from two Italian medical research universities. They wove it together in the order that symptoms appeared to trace the evolution of sepsis from subtle symptoms to critical stages. This structure, Herrero says, provided an ideal opportunity for an AI model to detect sepsis at different junctures in patient care.
The machine learning tool recognizes the symptoms of a person who may be developing sepsis, Herrero says.
“We get better performance than a regular doctor,” he says.
